Tencent Opens Hy4: A 770-Billion-Parameter AI Model Reshaping Global Tech and Real Estate Demand

Tencent has released a preview of Hy4, its largest open-source language model to date, built with 770 billion parameters and a context window exceeding 1 million tokens. Launched on August 28, 2026, the model is a direct challenge to Meta, Google and OpenAI in a market where pricing for AI services could shift dramatically before year end.

Hy4 targets three core use cases: coding, office automation and scientific research. Tencent is positioning it as a frontier-level open model while offering free access for two weeks through its WorkBuddy and CodeBuddy platforms.

Quick Answer

• Hy4 has 770 billion parameters in total architecture, with only 49 billion active during inference (a Mixture of Experts design)

• The context window exceeds 1 million tokens, enough to process documents hundreds of pages long in a single request

• API access via Tencent Cloud TokenHub and OpenRouter costs roughly $0.834 per 1K input tokens and $2.501 per 1K output tokens

• Cached tokens cost just $0.042 per 1 million, an aggressively low rate for repeat queries

• Free access through WorkBuddy and CodeBuddy lasts 14 days; the previous Hy3 model stays free until September 30, 2026

• In internal blind testing, Hy4 Preview reportedly outperformed Z.ai GLM-5.3 and Moonshot Kimi K3, two leading Chinese rivals

Key Facts

Release date: August 28, 2026. Tencent adopted a 'preview-first' strategy, releasing an open preview ahead of the final version to gather developer feedback before full launch.

Model scale. With 770 billion total parameters and only 49 billion active per token, Hy4 uses a Mixture of Experts (MoE) architecture that reduces computing costs by activating just a fraction of parameters per request. For comparison, Meta's Llama 3.1 dense model contains 405 billion parameters.

Pricing structure. Input tokens cost $0.834 per 1K and output tokens $2.501 per 1K, positioning Hy4 as enterprise-accessible but not the cheapest option on the market. Its standout advantage is caching at just $0.042 per 1 million tokens.

Access channels. The model is integrated into Tencent's own products, including WorkBuddy, CodeBuddy, Yuanbao and ima. The API is also open through Tencent Cloud TokenHub and OpenRouter, giving developers outside China straightforward access.

Technical improvements. According to Tencent, Hy4 uses technologies such as Gated DeepSeek Sparse Attention and IndexCache to make long-context processing cheaper and to reuse sparse indexes across layers, alongside a significant boost in training data volume and quality compared with Hy3.

Competitive benchmarks. In internal blind tests, Hy4 Preview reportedly beat Z.ai GLM-5.3 and Moonshot Kimi K3, while showing mixed results against Alibaba's Qwen 3.8 Max in third-party benchmarks, according to industry coverage of the release.

Open-source strategy. Releasing the weights of a 770-billion-parameter model on platforms like Hugging Face, ModelScope, GitCode and CNB is a direct statement in the enterprise AI market currently dominated by closed models from OpenAI and Google. Tencent is betting on mass adoption to build dependency on its cloud infrastructure.

Roadmap. Tencent has confirmed that additional models in the Hy4 series will follow 'in the near future,' signaling rapid iteration.

FAQ

What is Tencent Hy4 and how does it differ from the previous version?

Hy4 is the fourth-generation large language model from Tencent's Hunyuan division. Unlike Hy3, it contains 770 billion parameters (49 billion active) and its context window has grown to over 1 million tokens. The model is optimized for coding, office productivity and scientific research.

How much does it cost to use Hy4 via API?

Approximately $0.834 per 1,000 input tokens, $2.501 per 1,000 output tokens, and $0.042 per 1 million cached tokens, making cached inference dramatically cheaper than standard processing.

Is Hy4 available for free?

Yes, for a limited time. Through WorkBuddy and CodeBuddy, the model is free for 14 days after release. The previous Hy3 model remains free until September 30, 2026.

Can developers outside China use Hy4?

Yes. The API is available through OpenRouter, giving global developers access, along with direct access via Tencent Cloud TokenHub.

Why is Tencent open-sourcing such a large model?

Open weights encourage mass adoption and build reliance on Tencent Cloud infrastructure. It is a strategic move to capture market share in enterprise AI, particularly against closed offerings from OpenAI, Google and Anthropic.

What tasks does Hy4 handle best?

According to Tencent, three areas stand out: automating code writing and refactoring, processing large-volume office documents (thanks to the 1 million+ token context window), and scientific research involving analysis of lengthy texts.

When will the full version of Hy4 launch?

No exact date has been announced. Tencent is following a 'preview-first' approach and has promised further models in the Hy4 series in the coming months.

How does this affect the broader AI market?

Releasing an open-source model with 770 billion parameters puts downward pressure on pricing for closed models. Competition between Tencent, Meta (Llama) and Alibaba (Qwen) is accelerating the decline in AI inference costs for businesses worldwide.

The race among Chinese tech giants for open-source AI leadership indirectly shapes Southeast Asia's appeal as a technology hub. Thailand is actively courting data centers and AI startups, a trend that is fueling additional demand for commercial and residential property in Phuket and Bangkok, particularly among remote-working IT professionals relocating to the region.

Source: Tencent

Ready to invest in Thailand? Our experts will help you find the perfect property.